duskcode

duskcode is a terminal coding agent written in Rust. It connects to model APIs and provides tools for reading and editing project files, running shell commands, and delegating tasks to specialist agents. Sessions are stored locally in SQLite.

Install

curl -fsSL https://raw.githubusercontent.com/DroBov1Tya/dusk_code/main/install.sh | sh

The installer asks what to install: the terminal client duskcode, the desktop client duskcode-gui, or both. DUSKCODE_INSTALL=tui, gui or both answers ahead, for a server or a script:

curl -fsSL https://raw.githubusercontent.com/DroBov1Tya/dusk_code/main/install.sh | DUSKCODE_INSTALL=tui sh

Each part has its own script, which finds an existing install and updates it in place:

• scripts/install-tui.sh downloads a release build for macOS arm64 or Linux x86_64 / arm64, verifies its checksum, and links duskcode into /usr/local/bin or ~/.local/bin.
• scripts/install-gui.sh builds the desktop client from a checkout (or downloads it once releases carry it), puts it beside duskcode, and adds Duskcode.app to ~/Applications on macOS or a desktop entry on Linux. The build needs a recent Rust: rustup update stable.

Run the terminal client from a project folder:

duskcode

Use /connect to configure a provider and /model to select a model. The supported protocols are OpenAI-compatible Chat Completions and Anthropic Messages.

To build from source:

cargo build --release
./target/release/duskcode

The desktop client is a separate binary, so a server never builds it:

cargo build --release -p dusk_gui
./target/release/duskcode-gui

Features

• MAIN, ORCHESTRATOR, and PLAN modes for direct work, delegation, and planning.
• Specialist agent definitions, skills, modes, and prompts stored as editable Markdown.
• MCP server connections and configurable tools and agents through /settings.
• Plugins that bring skills, agents, MCP tools, providers, hooks and platforms such as the built-in Telegram bot, found at runtime and switched on in /plugins; see docs/plugins.md.
• A transcript of responses, reasoning, and tool calls, with context and token usage.
• Background shell commands with status and output in /jobs.
• Task tracking, plus goals with acceptance criteria through /goal.
• Saved chats through /resume, context summaries through /compact, and new threads from a summary through /fork.
• Configuration and chat export through /migrate; archives include API keys.
• Release updates through /update, with stable and dev channels.

Architecture

duskcode TUI
     │
 dusk_core
     ├── Model providers: OpenAI-compatible / Anthropic
     ├── Agent runtime and tools: files / shell / MCP
     ├── Plugins: skills / agents / MCP / providers / hooks / platforms
     └── Local state and configuration: SQLite / YAML / Markdown

Component	Responsibility
projects/core	Configuration, providers, agents, tools, runtime, and state
projects/tui	Terminal interface, input, navigation, and images
projects/external/telegram	The Telegram bot of the built-in telegram plugin
projects/evals	Evaluation runner for coding tasks
resources/	Bundled modes, agents, skills, plugins, and prompts

Configuration and data

Location	Contents
~/.config/duskcode/config.yml	Providers, model, limits, tools, agents, skills, MCP servers, appearance, and logging
~/.config/duskcode/rules/	Agents, skills, modes, prompts, plugins, and user memory
~/.config/duskcode/plugin-data/	What plugins keep, such as a paired Telegram owner
~/.config/duskcode/session/	Session database and artifacts
~/.config/duskcode/logs/	Application logs
~/.config/duskcode/mcp-auth/	MCP OAuth tokens
~/Documents/duskcode/	Chat workspaces and files; projects/ holds project memory
~/.local/share/duskcode/	Default installation directory; the installer also supports XDG_DATA_HOME

DUSKCODE_LOG sets the log level. Logging can be disabled with logging.enabled: false; optional OpenTelemetry tracing is configured under logging.traces. DUSKCODE_GRAPHICS=kitty|iterm2|sixel selects a terminal image protocol. duskcode remove starts the interactive uninstaller.

Limitations

• The project is pre-1.0; commands, configuration, and database schema may change.
• Shell commands run with the user's permissions. The approval setting controls confirmation prompts; it does not provide an operating-system sandbox.
• Model behavior and API compatibility depend on the configured provider and model. Requests sent to external providers and MCP servers are subject to their data policies.
• Release builds target Linux and macOS; there is no Windows build.
• Image display depends on terminal graphics support, with a character preview fallback.

License

MIT. Bundled skills retain their own licenses; see resources/skills/NOTICE.md.